@astermind/astermind-elm/dist/tasks/EncoderELM.d.ts

JavaScript Extreme Learning Machine (ELM) library for browser and Node.js.

import { ELM } from '../core/ELM';
import { ELMConfig, Activation } from '../core/ELMConfig';
export declare class EncoderELM {
    elm: ELM;
    private config;
    private online?;
    private onlineInputDim?;
    private onlineOutputDim?;
    constructor(config: ELMConfig);
    /** Batch training for string → dense vector mapping. */
    train(inputStrings: string[], targetVectors: number[][]): void;
    /** Encode a string into a dense feature vector using the trained model. */
    encode(text: string): number[];
    /**
     * Begin an online OS-ELM run for string→vector encoding.
     * Provide outputDim and either inputDim OR a sampleText we can encode to infer inputDim.
     */
    beginOnline(opts: {
        outputDim: number;
        inputDim?: number;
        sampleText?: string;
        hiddenUnits?: number;
        ridgeLambda?: number;
        activation?: Activation;
        weightInit?: 'uniform' | 'xavier' | 'he';
        forgettingFactor?: number;
        seed?: number;
    }): void;
    /**
     * Online partial fit with *pre-encoded* numeric vectors.
     * If not initialized, this call seeds the model via `init`, else it performs an `update`.
     */
    partialTrainOnlineVectors(batch: Array<{
        x: number[];
        y: number[];
    }>): void;
    /**
     * Online partial fit with raw texts and dense numeric targets.
     * Texts are encoded + normalized internally.
     */
    partialTrainOnlineTexts(batch: Array<{
        text: string;
        target: number[];
    }>): void;
    /**
     * Finalize the online run by publishing learned weights into the standard ELM model.
     * After this, the normal encode() path works unchanged.
     */
    endOnline(): void;
    loadModelFromJSON(json: string): void;
    saveModelAsJSONFile(filename?: string): void;
}